* Python: Add AgentLoopMiddleware for re-running agents in a loop
Add `AgentLoopMiddleware`, an `AgentMiddleware` that re-runs the wrapped
agent in a loop. A single configurable class covers three common patterns,
each with a convenience classmethod factory:
- Ralph loop (`.ralph(...)`): no exit criteria, with feedback tracking
(`record_feedback`/`progress`), progress injection (`inject_progress`),
optional fresh context per iteration (`fresh_context`), and an early-stop
completion signal (`is_complete`).
- Predicate (`.with_predicate(...)`): loop while a `should_continue` callable
returns True (e.g. paired with `todos_remaining`/`background_tasks_running`).
- Judge (`.with_judge(...)`): a second chat client decides whether the original
request was answered, using a `JudgeVerdict` structured-output response.
The loop also auto-resolves pending function-approval / user-input requests via
an `on_approval_request` callable (bounded by `max_approval_rounds`), and the
next iteration's input is controlled by `next_message`. Supports both streaming
and non-streaming runs.
Exports `AgentLoopMiddleware`, `JudgeVerdict`, `todos_remaining`, and
`background_tasks_running`. Adds tests, a sample, and docs.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Python: Refine AgentLoopMiddleware API and sample
- with_judge: add criteria list with {{criteria}} templating into judge
instructions plus an agent-side instruction; add fresh_context, additional
judge feedback relay; default judge max_iterations.
- should_continue is now required and positional; supports (bool, str|None)
feedback tuples surfaced to next_message/record_feedback via feedback kwarg.
- Judge forwards full multi-modal request and response messages.
- Default max_iterations=10 (explicit None = unbounded); removed is_complete and
Ralph terminology; ShouldContinueResult is a real TypeAlias.
- Sample: stream all loops, print iteration counts via injected user-block
boundaries (robust to function calling), <role>: content formatting, per-method
expected output, and a looping todo sample.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Python: Fix CI checks for AgentLoopMiddleware
- Resolve pyright errors in _loop.py: drop the always-true final_result None
check (the while loop always assigns it) and cast finish_reason to the
AgentResponse constructor's expected type.
- Apply pyupgrade --py310-plus: import TypeAlias from typing.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Python: Resolve mypy/pyright disagreement on finish_reason
pyright infers AgentResponse.finish_reason as including str and rejects the
direct assignment, while mypy considers a cast redundant. Drop the cast and
suppress only pyright with a targeted reportArgumentType ignore, satisfying
both type checkers.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Python: Add todo+judge AgentLoopMiddleware sample
Add a second AgentLoopMiddleware sample that composes two criteria in one
should_continue predicate: a TodoProvider check (evaluated first) and a
report-style judge chat client (evaluated once todos are complete) that grades
the assembled report against shared requirements. Register it in the middleware
samples README.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Python: Compose todo+judge loops as two middleware
Rework the todo+judge sample to compose two AgentLoopMiddleware on the agent
itself (middleware=[judge_loop, todo_loop]) instead of a single hand-written
predicate. The inner todos_remaining loop drafts the report todo-by-todo and the
outer with_judge loop re-runs it until an editor chat client judges the report
publication-ready, reusing the built-in helpers.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Reset session for fresh_context loops via snapshot/restore
AgentLoopMiddleware.fresh_context previously only reset context.messages,
so with an attached session each iteration still reloaded the local
transcript or re-threaded the service-side conversation id and the model
saw the accumulated history. Snapshot the session once before the loop
(via to_dict) and restore it (from_dict + field copy) between iterations,
so every pass starts from the pre-loop baseline. The final iteration's
pass is persisted (no restore after the terminating iteration), so a
subsequent agent.run continues from there.
Removed the obsolete warning, updated docstrings and core AGENTS.md, and
added tests: a snapshot/restore round-trip, a session-reset
streaming x fresh_context x inject_progress x store matrix across multiple
runs and loop iterations, and response_format parsing across the loop.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Updated samples and docstrings
---------
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Integrate shell tool into AgentHarness
* Validate shell_executor exposes as_function() with a clear TypeError
Addresses PR review feedback: a public factory should fail fast with an
actionable error rather than a cryptic AttributeError when an incompatible
shell_executor is supplied. Validation happens upfront, regardless of whether
the client supports shell tools.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Type shell harness params via TYPE_CHECKING import
Addresses PR review feedback: type shell_executor and
shell_environment_provider_options instead of Any, using a TYPE_CHECKING
import from agent_framework_tools.shell. The import never executes at
runtime, so there is no circular dependency, and the lazy runtime import of
ShellEnvironmentProvider is retained. Since ShellExecutor is a protocol
without as_function(), the validated getattr result is invoked directly.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
---------
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Fix MCP allowed_tools empty list handling
When allowed_tools is set to an empty list [], the falsy check
'if not self.allowed_tools' incorrectly treats it as unconfigured
(same as None), causing all tools to be exposed. Change to an
explicit 'is None' check so that an empty list correctly results
in no tools being allowed.
Co-authored-by: Azure SRE Agent <noreply@microsoft.com>
* Clarify allowed_tools docstring: None vs [] semantics
Per Eduard's review on PR #6296: explicitly document that None exposes all tools and [] exposes none, across all four MCPTool / MCPStdioTool / MCPStreamableHTTPTool / MCPWebsocketTool docstrings.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* allowed_tools docstring: recommend load_tools=False for full disable
Per Eduard's follow-up on PR #6296: `load_tools=False` is the cleaner idiom when you don't want to expose any tools. Reframe `allowed_tools=[]` in the docstring as a runtime guard / inspection-only path and cross-reference `load_tools`.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
---------
Co-authored-by: Azure SRE Agent <noreply@microsoft.com>
Co-authored-by: Giles Odigwe <79032838+giles17@users.noreply.github.com>
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Parse structuredContent from MCP CallToolResult (#3313)
The _parse_tool_result_from_mcp method only iterated over the content
field from CallToolResult, ignoring the structuredContent field entirely.
MCP servers that return JSON data via structuredContent (e.g., Power BI
MCP) appeared to return None.
Add handling for structuredContent: when present, serialize it as JSON
text and append it to the result list. This preserves the data for the
LLM while maintaining backward compatibility with existing behavior.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Python: Parse MCP CallToolResult.structuredContent field to prevent tool results returning None
Fixes#3313
* Address review feedback: add default=str to json.dumps and remove .checkpoints/
- Add default=str to json.dumps for structuredContent serialization so
non-JSON-serializable values (e.g. bytes) degrade gracefully instead
of raising TypeError
- Remove all .checkpoints/ runtime artifacts from the repository
- Add **/.checkpoints/ to .gitignore to prevent future accidental commits
- Add test for non-serializable structuredContent values
Fixes#3313
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Address review feedback for #3313: Python: MCP CallToolResult.structuredContent field is not parsed, causing tool results to return None
---------
Co-authored-by: Copilot <copilot@github.com>
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Add sampling guardrails to MCP tools
Add approval, token, and request-count controls to the MCP sampling
callback used when an MCPTool is configured with a chat client.
- Add `sampling_approval_callback`, `sampling_max_tokens`, and
`sampling_max_requests` parameters to `MCPTool` and its
`MCPStdioTool`, `MCPStreamableHTTPTool`, and `MCPWebsocketTool`
subclasses, positioned directly after `client`.
- Gate each server-initiated `sampling/createMessage` request behind the
approval callback, which denies by default when no callback is provided.
- Clamp the requested `maxTokens` to `sampling_max_tokens` and enforce a
per-session request count via `sampling_max_requests`.
- Log incoming sampling requests at WARNING level (counts only).
- Export `SamplingApprovalCallback` from the public API.
- Add tests, a sample, and documentation updates.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Make sampling denial message context-aware
Distinguish the deny-by-default case (no approval callback configured)
from an explicit denial by a configured `sampling_approval_callback`, so
the returned ErrorData message is accurate for callback-driven denials
and exceptions.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
---------
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Filter MCP tool kwargs to declared params via allowlist
Previously MCPTool combined framework runtime kwargs (from
FunctionInvocationContext.kwargs) with the LLM-supplied arguments and
stripped only a hardcoded denylist of known framework keys before
forwarding to the MCP server. Any new framework-injected kwarg leaked to
the server unless the denylist was updated.
Switch to an allowlist built from each tool's declared parameters
(inputSchema.properties). Only declared params are forwarded; everything
else is stripped. Add an `additional_tool_argument_names` constructor
argument so users can opt extra names back in, globally (Sequence[str])
and/or per remote tool name (Mapping with reserved "*" global key). The
existing denylist is kept as a safety net for framework-named params a
server declares in its schema; explicitly opted-in extras always win. The
reserved _meta handling is unchanged.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Address MCP allowlist review comments and fix reload arg loss
- Fix pyright reportUnknownArgumentType in _load_tools (cast schema properties).
- Register declared param names before the existing-tool skip guard so that
tool-list reloads preserve the allowlist for already-loaded tools (previously
unchanged tools silently dropped all declared args after a background reload).
- Handle bare-string values in an additional_tool_argument_names mapping instead
of iterating their characters.
- Clarify the framework denylist comment: explicit extras override the denylist.
- Make the extras-override-denylist test unambiguous (opt in a denylisted name).
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
---------
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Fix per-service-call history persistence with server-storing clients
When an Agent set require_per_service_call_history_persistence=True together
with a HistoryProvider, and the chat client stored history server-side by
default (e.g. OpenAIChatClient, STORES_BY_DEFAULT=True), the external history
provider was silently never persisted.
Unify persistence on the per-service-call middleware: when the flag is set and
a HistoryProvider exists, the middleware is always installed and owns
persistence. service_stores_history now only selects middleware behavior:
- service does not store: load providers and drive the function loop with a
local sentinel conversation id, or
- service stores: skip loading (the service owns history) and persist each
service call while the real conversation id flows through.
Also rationalize chat-options handling in _prepare_run_context:
- _merge_options now skips None overrides and strips remaining None values, so
an unset `store` is never forwarded and the service decides its own default.
- Resolve `store` and `conversation_id` once from a single combined view
(effective_options) instead of probing both default and runtime dicts; the
auto-injection and per-service-call resolution now agree on conversation_id.
Fixes#5798
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Correct as_agent() docstring: persistence is per service call, not once per run
Address PR review: when the client stores history server-side, the
per-service-call middleware still persists after each model call; only
provider loading is skipped. The previous "persist once per run()" wording
contradicted the implementation.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Address PR review: docs, missing-conversation-id warning, and tests
- Clarify that require_per_service_call_history_persistence is a no-op when no
HistoryProvider is present (docstrings in _agents.py and _clients.py).
- Warn on every service call when the client stores history server-side but
returns no conversation_id, so the (uncommon) loss of cross-turn resumability
cannot fail silently.
- Add tests: storing client + existing conversation_id does not raise and the id
propagates; two runs on the same session keep persisting with a stable
service_session_id and no provider loading; storing-without-conversation-id
warns per call.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
---------
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* MCP long-running task support in Python
* Fix pyupgrade and AGENTS.md reconnect description
- pyupgrade: drop forward-reference string annotations in _mcp.py (Python 3.10+ resolves them natively now that MCPTaskOptions is defined before use).
- AGENTS.md: align reconnect description with current behavior. Phase 1 (initial tools/call) does NOT retry on connection loss; raises 'connection lost; task state unknown' instead, so a server that accepted the request but lost the response cannot start the operation twice. Phase 2 (tasks/get / tasks/result) still reconnects once against the same task_id.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Fix bandit nosec marker for CI pipeline
* Address PR feedbacks
* Clarifiied comments and addressed more PR feedbacks.
---------
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Fix compaction message-id collisions and tool-loop summary persistence
Fixes two bugs in the compaction strategies:
- #5237: incremental group annotation assigned message ids by position
within the re-annotated slice, so moving the re-annotation start back to
a previous group start restarted ids at 0 and produced collisions
(e.g. a user message reusing an assistant message's id), merging groups
and causing tool-result compaction to wrongly exclude messages.
group_messages/_ensure_message_ids now take an id_offset and guard
against existing-id collisions; annotate_message_groups threads the
slice start index through as the offset.
- #4991: the function-invocation loop copied the message list each
iteration, so summaries inserted by compaction landed in a throwaway
copy and were lost across tool-loop iterations (only the persistent
excluded flags survived). _prepare_messages_for_model_call now compacts
the list in place when messages is a list, so inserted summaries persist.
Adds regression tests (incremental id uniqueness, existing-id collision
avoidance, idempotency, and tool-loop summary persistence including
streaming and conversation-id modes).
Also adds a summarization.py sample demonstrating SummarizationStrategy
directly with a real client, and reworks advanced.py with tool-call
groups and a real summarizer.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Guard incremental message-id assignment against prefix-id collisions
Addresses PR review on #5237: _ensure_message_ids only guarded against
collisions within the re-annotated slice. A preexisting (e.g. user-supplied)
id in the preserved prefix could still be reassigned in the suffix when the
id was numerically out of position, merging groups across the re-annotation
boundary again.
group_messages/_ensure_message_ids now accept reserved_ids, and
annotate_message_groups passes the preserved prefix's ids so auto-assigned
suffix ids never collide across the full list. Adds a regression test
reproducing the out-of-position prefix-id collision.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
---------
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Add MCP-based skills discovery (McpSkill, McpSkillsSource, McpSkillResource)
Implement Agent Skills discovery over MCP following the SEP-2640 convention:
- McpSkillsSource: reads skill://index.json to discover skills served by an MCP server
- McpSkill: lazily fetches SKILL.md content via resources/read on demand
- McpSkillResource: wraps MCP resource results (text and binary)
- Path traversal protection in get_resource for defense in depth
- Samples for Foundry Toolbox and standalone MCP skills server
- Comprehensive unit tests (514 lines)
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Address PR review comments: rename to MCP* convention, fix error handling and samples
- Rename McpSkill/McpSkillResource/McpSkillsSource to MCPSkill/MCPSkillResource/MCPSkillsSource
- Add data-URI prefix stripping for blob resource decoding
- Let non-McpError exceptions propagate from get_resource()
- Fix contradictory test comment
- Use interactive input() in mcp_based_skill sample
- Remove misleading sample output block
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Restore debug logging for McpError in get_resource()
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Use AzureCliCredential in Foundry toolbox skills sample for consistency
Replace DefaultAzureCredential with AzureCliCredential to match the
credential convention used in all other samples.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Use MCPStreamableHTTPTool in MCP skills sample
Replace raw mcp library imports (ClientSession, streamable_http_client)
with the framework's MCPStreamableHTTPTool to keep MCP server connections
consistent regardless of whether skills are enabled.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Branch on McpError.error.code so only not-found errors return empty
Previously _try_read_index() and get_resource() swallowed every McpError
as 'no skills available', making auth failures, server crashes, and
connection drops indistinguishable from a server that simply has no
skills.
Now only two codes are treated as not-found:
- -32002 (MCP-spec Resource not found)
- -32601 (METHOD_NOT_FOUND — server lacks resources/read)
All other McpError codes and non-McpError exceptions propagate with a
warning log, surfacing real failures visibly.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Add tests for non-McpError and non-not-found error propagation in MCP skills
Cover the re-raise branch in MCPSkill.get_resource for plain
ConnectionError/TimeoutError, the generic McpError (code 0) propagation
on get_resource, and TimeoutError propagation in _try_read_index.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Revert "Use MCPStreamableHTTPTool in MCP skills sample"
This reverts commit f31ed0ded9.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Introduce MCP_SKILLS experimental feature for MCP skill classes
Add a separate MCP_SKILLS feature ID to ExperimentalFeature enum and
use it for MCPSkillResource, MCPSkill, and MCPSkillsSource, since their
promotion timeline is partly outside of our control.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
---------
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Python: progressive tool exposure via FunctionInvocationContext
Add first-class progressive tool exposure to the Python core function-calling
loop. Tools can now add or remove real FunctionTool schemas at runtime via the
injected FunctionInvocationContext, taking effect on the next iteration of the
loop.
- FunctionInvocationContext gains a live `tools` list plus experimental
`add_tools()` / `remove_tools()` helpers (feature: PROGRESSIVE_TOOLS).
- The function-calling loop establishes a run-local, normalized tools list and
threads it into the context at both invocation paths so mutations propagate.
- Add a sample (dynamic_tool_exposure.py) and a tools samples README, including
a note that CodeAct providers (Monty/Hyperlight) use their own provider-level
tool management instead.
Supersedes #3877.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Validate non-negative input in dynamic_tool_exposure sample tools
Address review feedback: factorial and fibonacci now return an error
message for negative n instead of producing incorrect results.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Make add_tools atomic and surface swallowed function errors
Address review feedback on progressive tool exposure:
- add_tools now validates the full batch against a throwaway copy before
committing, so a duplicate-name clash partway through a sequence leaves
the live tool list unchanged (all-or-nothing).
- _auto_invoke_function now logs a warning (with traceback) when a tool
raises, so contract errors such as a duplicate-name ValueError from
add_tools are debuggable without enabling include_detailed_errors.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Avoid retaining tracebacks when logging swallowed function errors
Logging with exc_info=exc fed the exception traceback to the logging
machinery, whose frame references created reference cycles collected
lazily by the cyclic GC. On Windows that could drop a hyperlight
WasmSandbox on a non-owning thread ("unsendable, dropped on another
thread"), crashing the xdist worker. Log a pre-formatted message with
the exception repr instead, so no traceback object is retained.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* added missing decorator
---------
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Python: Fix OTLP HTTP base-endpoint losing /v1/{signal} auto-append
Per the OTel spec, OTEL_EXPORTER_OTLP_ENDPOINT is a *base* URL for HTTP —
the SDK auto-appends /v1/traces, /v1/metrics, /v1/logs when it reads the
env var directly. Signal-specific endpoint env vars are *full* URLs used
verbatim.
_get_exporters_from_env read the base endpoint and forwarded it as the
constructor ``endpoint=`` argument, which the SDK always treats as a full
signal URL. As a result, with OTEL_EXPORTER_OTLP_ENDPOINT=http://localhost:4318
and HTTP protocol, the exporter sent to http://localhost:4318 instead of
http://localhost:4318/v1/traces (and likewise for metrics/logs).
Replicate the spec's auto-append here when falling back to the base
endpoint under HTTP. gRPC behavior is unchanged.
* Python: Fix mypy type errors in OTLP endpoint assignment
Pre-declare traces_endpoint, metrics_endpoint, logs_endpoint as
str | None before the if/else block. Mypy inferred str from the
if-branch f-string assignments and then rejected the str | None
expressions in the else-branch as incompatible.
* Python: feat(evals): RubricScore type + EvalScoreResult.dimensions
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Python: feat(foundry-evals): RubricDimension + GeneratedEvaluatorRef + accept in evaluators=
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Python: feat(evals): parse rubric_scores from output items + assertion helpers
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Python: feat(evals): BaseAgent.as_eval_source / Workflow.as_eval_source
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Python: feat(foundry-evals): EvalGenerationSource + generate_rubric helper
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Python: feat(foundry-evals): YAML config loader + sample
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Python: fix(evals): address PR review feedback
Addresses 4 Copilot review comments on PR #6101:
1. assert_dimension_score_at_least: drop the (not evaluator or found_any) guard so require_applicable=True correctly raises when the named evaluator produces no entries for the dimension. Adds TestRubricAssertions covering the regression.
2. GeneratedEvaluatorRef docstring: reword to describe actual behaviour (pinning recommended, not required) so it matches the dataclass default and FoundryEvals warning path.
3. _poll_generation_job: switch from asyncio.get_event_loop() to get_running_loop() and bound the per-iteration sleep by remaining time, matching _poll_eval_run.
4. generate_rubric: type category as Literal['quality','safety'] and validate at the entry point with a ValueError; drop the silent 'invalid -> quality' rewrite in _generation_job_to_ref. Adds a regression test.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Python: feat(foundry-evals): hosted-agent-aware rubric generation
* Auto-detect hosted Foundry agents in agent_as_eval_source: when the
agent's chat_client exposes a string agent_name (the convention used
by RawFoundryAgentChatClient for PromptAgents/HostedAgents), emit a
type='agent' EvalGenerationSource so the service fetches instructions
and tools from the agent registry instead of relying on the local
wrapper (which holds neither for hosted agents).
* Add hosted_agent_version kwarg and a new agent_version field on
EvalGenerationSource so PromptAgent runs can pin to a specific hosted
version for reproducible rubric generation.
* Add force_prompt_source escape hatch to bypass auto-detection and
always emit a rendered prompt dossier - useful when the local wrapper
carries overrides the service-side agent doesnt see.
* Fix _to_sdk_source for dataset sources: SDK ctor takes name=/version=,
not dataset_name=/dataset_version=. The mismatch would raise TypeError
against the real azure-ai-projects 2.3.0a* SDK; only unmocked
integration paths were affected.
Tests cover: auto-detection happy path, versionless hosted agent,
explicit hosted_agent_version forwarding, force_prompt_source override,
non-string chat_client attrs (MagicMock test doubles) not mis-detected,
agent_version forwarded through _to_sdk_source, and the corrected
dataset SDK kwarg names.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* fix(foundry-evals): accept canonical dimension_scores key per docs
The published Foundry rubric-evaluator output (Microsoft Learn 'Rubric evaluators' reference) places per-dimension breakdowns under properties.dimension_scores, not properties.rubric_scores. The parser now tries dimension_scores first and falls back to rubric_scores for preview-build compatibility, and tolerates non-list payloads (e.g. MagicMock auto-attrs) by trying the next candidate when parsing yields zero entries.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* feat(foundry-evals): add manual create_rubric_evaluator
Adds FoundryEvals.create_rubric_evaluator as the agent-framework surface over project_client.beta.evaluators.create_version. This is the manual counterpart to generate_rubric: callers supply RubricDimension instances (authored locally, ported from another framework, or hand-tuned) and we POST a RubricBasedEvaluatorDefinition. The service auto-attaches the non-editable residual dimension (general_quality for quality, general_policy_compliance for safety).
Per the Microsoft Learn 'Rubric evaluators' reference, the auto-generation path (create_generation_job) is primarily a portal/UI feature; external SDK clients with rich local agent context are better served by manual create_version. This keeps generate_rubric for users who want to round-trip through a Foundry-registered agent.
Validation up front: weight must be in [1,10], ids unique, descriptions non-empty, pass_threshold in [0,1]. The returned GeneratedEvaluatorRef is identical in shape to one obtained from generate_rubric, so downstream evaluators= lists work unchanged.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* samples(foundry-evals): manual rubric sample + namespace re-exports
Adds evaluate_with_manual_rubric_sample.py demonstrating the end-to-end dev scenario for FoundryEvals.create_rubric_evaluator: hand-author a list of RubricDimension, register via create_rubric_evaluator, then use the pinned GeneratedEvaluatorRef alongside built-in evaluators in an agent regression run.
Also re-exports RubricDimension, GeneratedEvaluatorRef, build_sources, and load_evals_config from agent_framework.foundry (both the lazy runtime shim and the type stub) so the rubric samples can import everything from a single namespace; the auto-generate sample was previously broken because the shim was missing build_sources / load_evals_config.
Updates the foundry-evals README with a chooser entry for the two rubric paths.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* feat(foundry-evals): remove rubric creation flows; keep consumption only
Reframes agent-framework as a pure consumer of Foundry rubric evaluators: scoring against rubrics that already exist (authored in the Foundry portal or via the dedicated SDK / REST surface) instead of creating them from the SDK.
Removed creation surface area:
- FoundryEvals.generate_rubric (auto-generate path) and create_rubric_evaluator (manual path), plus all _GenerationSdkTypes / _ManualRubricSdkTypes / _to_sdk_dimensions / _coalesce_generation_sources / _to_sdk_source / _poll_generation_job / _generation_job_to_ref / _evaluator_version_to_ref / _get_beta_evaluators / _import_*_sdk_types helpers.
- EvalGenerationSource (the input source discriminator), RubricDimension (the input dimension type), agent_as_eval_source / workflow_as_eval_source / _detect_hosted_foundry_agent helpers, and the YAML-config loader (_evals_config.py with RubricGenerationSpec / RubricSourceSpec / parse_evals_config / load_evals_config / build_sources).
- BaseAgent.as_eval_source / Workflow.as_eval_source plus the _render_agent_dossier / _render_workflow_dossier helpers in core. These existed only to feed the now-removed generation pipeline.
- Samples evaluate_with_generated_rubric_sample.py, evaluate_with_manual_rubric_sample.py, and evaluators.yaml. Replaced with a short README section showing how to reference an existing rubric evaluator via GeneratedEvaluatorRef.
Kept (consumption surface):
- GeneratedEvaluatorRef, slimmed to (name, version, display_name). Still accepted alongside built-in evaluator strings in FoundryEvals(evaluators=[...]). Versionless refs still warn.
- RubricScore on EvalScoreResult.dimensions plus EvalResults.assert_dimension_score_at_least for per-dimension CI gates.
- _parse_dimension_entries / _extract_rubric_scores output parsing (both canonical dimension_scores and the legacy rubric_scores key).
Tests: 160/160 foundry unit tests and 71/71 core local-eval tests pass; pyright is clean across changed files. The pre-existing tests/core/test_telemetry.py::test_detect_hosted_fallback_import_error failure is unrelated and reproduces on the prior commit.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* samples(foundry-evals): add evaluate_with_rubric_sample
Adds a runnable end-to-end sample showing how to consume a pre-existing rubric evaluator created in Foundry: reference it with GeneratedEvaluatorRef(name, version), mix it with built-in evaluators in FoundryEvals, and gate CI with assert_dimension_score_at_least on a specific dimension.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* fix(foundry-evals): satisfy mypy on _fetch_output_items
mypy infers OutputItemListResponse.sample as dict[str, object] | None while pyright correctly infers the typed Sample model. Cast to Any so both type checkers accept the attribute access pattern, rename the local to avoid shadowing the inner-loop sample binding, and drop the now-stale pyright suppressions.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* docs(foundry-evals): drop unpublished rubric-evaluators learn.microsoft.com link
The Adaptive Evals authoring docs are not yet published on Microsoft Learn, so the link 404s. Keep the descriptive text without the broken hyperlink; we can re-add it once the docs ship.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* test(foundry-evals): hoist repeated local imports to module top
Per code review feedback (eavanvalkenburg): the test file repeated 'from agent_framework_foundry._foundry_evals import ...' inside 22 test bodies and 'from agent_framework_foundry import GeneratedEvaluatorRef' inside 8 more. Move all of them to the existing top-level imports; the symbols are the same across tests and the local imports were redundant.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
---------
Co-authored-by: Ben Thomas <25218250+alliscode@users.noreply.github.com>
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* fix: safely serialize function-call arguments in core observability
Apply make_json_safe() to content.arguments in _to_otel_part() before
building the otel message dict, so that dataclass/framework payloads
(e.g. workflow request_info events) do not cause a TypeError when
_capture_messages() calls json.dumps().
Lift make_json_safe() into agent_framework._serialization (no new
external deps — dataclasses/datetime only) so the core observability
path can use it without a dependency on the ag-ui adapter.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* fix(core): safely serialize workflow request_info payloads in observability (#5733)
- Add make_json_safe() helper to recursively convert non-serializable objects
- Use make_json_safe() in _to_otel_part() for function_call arguments
- Fix CustomPayload test class to use @dataclass (resolves B903 lint error)
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* fix(serialization): guard callability and normalize dict keys in make_json_safe (#5733)
- Use callable(getattr(obj, method, None)) instead of hasattr() so that
non-callable attributes named model_dump/to_dict/dict do not raise
TypeError at runtime.
- Wrap each call in try/except TypeError to handle callables with
mandatory arguments gracefully.
- Convert dict keys to str() so that non-string keys (e.g. datetime,
int) cannot cause json.dumps to raise TypeError.
- Add regression tests for both scenarios.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Address observability serialization review feedback
---------
Co-authored-by: Copilot <copilot@github.com>
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Adding AgentFileStore and FileAccessProvider to support file ased operations for agents.
* Address PR review feedback on FileAccessProvider
- Probe symlinks on the unresolved candidate path so in-root symlinks
cannot silently pass and out-of-root symlinks surface the correct
error message.
- Validate matching_lines elements in FileSearchResult.from_dict and
raise a clean ValueError for non-mapping entries.
- Cap search regex pattern length (256 chars) via a new
_compile_search_regex helper to mitigate ReDoS, and surface the cap
in the file_access_search_files tool description.
- Skip non-UTF-8 files during filesystem search instead of aborting
the entire directory walk.
- Replace the module-scope trailing string in the data-processing
sample with comments to avoid Ruff B018.
- Remove the checked-in working/region_totals.md sample artifact so
the save flow works from a clean checkout.
- Expand the Windows stdout reconfiguration comment in task_runner.py
for clarity.
- Add tests for invalid/oversize regex, non-UTF-8 file search, and
in-root symlink rejection.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Fix mypy redundant-cast in FileSearchResult.from_dict
Use cast(list[object], ...) instead of cast(list[Any], ...) so the
cast represents a real type change (lists are invariant) and is no
longer flagged by mypy as redundant, while still satisfying pyright's
reportUnknownVariableType. Matches the existing pattern in _memory.py.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Tighten path normalization and directory resolution in FileAccess
- _normalize_relative_path now strips surrounding whitespace up front
so leading/trailing spaces never leak into file segments, and
rejects trailing path separators for file paths so 'foo/' is no
longer silently coerced to 'foo'.
- FileSystemAgentFileStore._resolve_safe_directory_path normalizes
with is_directory=True and maps an empty normalized result to the
root. This matches InMemoryAgentFileStore so whitespace-only
directory inputs resolve to the root instead of raising.
- Added tests for whitespace stripping, trailing-separator rejection,
and whitespace-only directory listing on the filesystem store.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Harden FileAccess search and atomic save in store API
- Add wall-clock timeout (10s) around regex scans so a pathological pattern (e.g. `(a+)+`) below the length cap cannot stall the event loop.
- Offload the InMemoryAgentFileStore regex scan to a worker thread, matching the filesystem store.
- Fail closed when `Path.is_symlink` raises during the safe-path probe so a permission error cannot silently bypass the symlink/reparse-point rejection.
- Add `overwrite: bool = True` to `AgentFileStore.write_file`; the in-memory store performs the check under the existing lock and the filesystem store uses `open(mode='x')` so concurrent callers cannot race past `overwrite=False`.
- `file_access_save_file` now relies on the atomic store call instead of a separate `file_exists` round-trip.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Fix Python 3.10 timeout handling and add directory arg to list/search tools
- Catch asyncio.TimeoutError in _run_search_with_timeout. In Python 3.10
asyncio.wait_for raises asyncio.exceptions.TimeoutError, which is
distinct from the builtin TimeoutError (the two were unified in 3.11).
Catching the asyncio alias works on every supported version.
- Add an optional directory parameter to file_access_list_files and
file_access_search_files so agents can enumerate / scope searches to
nested folders, not just the store root.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Address FileAccess review feedback: case, errors, signal, TOCTOU
- InMemoryAgentFileStore now stores (display_name, content) so list_files
and search_files return the original-case names callers wrote, matching
the behaviour of FileSystemAgentFileStore on case-preserving filesystems
and removing the silent in-memory vs. on-disk contract divergence.
- FileSystemAgentFileStore.read_file raises ValueError instead of letting
UnicodeDecodeError bubble for binary / non-UTF-8 input, restoring
symmetry with search_files (which still skips) and giving the tool
layer a recoverable type to translate.
- Tool wrappers now catch ValueError and OSError around every operation
and surface them as readable strings, so 'you used ..' and 'the file
already exists' are both reported to the model the same way instead of
the former crashing out as an unhandled exception.
- _search_files_sync logs per skipped non-UTF-8 file at WARNING and an
aggregate INFO summary so operators can distinguish 'no matches' from
'half the corpus was unreadable'.
- FileSystemAgentFileStore softens its docstrings to acknowledge the
inherent probe-then-open TOCTOU window. On POSIX both read and write
now pass O_NOFOLLOW so the kernel refuses if the leaf segment becomes
a symlink between the probe and the open. Windows has no equivalent
flag; the limitation is documented.
- Tests cover: case preservation on list/search, ValueError on non-UTF-8
read at the store and tool layer, tool-layer string responses for
path-traversal and oversized-regex inputs, search-skip log output,
symlink rejection on delete/search/list, and symlinked intermediate
directory rejection.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Address FileAccess nit comments: docstrings, enumerate, opt-in delete approval
- Expand FileSearchMatch/FileSearchResult.to_dict docstrings to explain why
the override is needed (__slots__ defeats the mixin's __dict__ iteration)
and why exclude/exclude_none are accepted-but-ignored (mixin signature
compatibility for callers like to_json).
- Use enumerate(lines, start=1) in _search_file_content so the +1 below is
no longer needed; rename loop variable to line_number for clarity.
- Add opt-in require_delete_approval: bool = False on FileAccessProvider.
When True, file_access_delete_file is registered with approval_mode
'always_require' so the host must approve every delete. Default False
preserves current behaviour and matches the .NET reference, but
deployments that want a safer-by-default posture can enable it.
- Add tests covering both delete approval modes.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* FileAccess: require delete approval by default
Flip the default for FileAccessProvider(require_delete_approval=...) from
False to True so destructive deletes are gated by host approval out of the
box. Callers that want the previous autonomous behaviour (which matches the
.NET reference) can pass require_delete_approval=False.
Tests updated accordingly.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Fixing linkinspector by installing Chrome for puppeteer first.
---------
Co-authored-by: Ben Thomas <25218250+alliscode@users.noreply.github.com>
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Align c# and python TodoProvider tool names
* Potential fix for pull request finding
Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>
* Address PR review: remove __slots__ and add typed schemas for tool params
- Remove __slots__ from TodoItem, TodoInput, and TodoCompleteInput classes
(not needed for low-instance-count objects and hinders dev scenarios)
- Add _TodoAddItemSchema and _TodoCompleteItemSchema TypedDicts to provide
proper JSON schema for todos_add and todos_complete tool parameters
- Use typing_extensions for Python 3.10 compatibility
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
---------
Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Add a HarnessAgent with available features and sample
* Fix formatting
* Address PR comments and fix mypy error
* Add web search support to HarnessAgent
* Fix build warning
* Apply suggestions from code review
Co-authored-by: Eduard van Valkenburg <eavanvalkenburg@users.noreply.github.com>
* Address PR comments
* Address PR comments
* Address further PR comments.
* Fix markdown broken link
---------
Co-authored-by: Eduard van Valkenburg <eavanvalkenburg@users.noreply.github.com>
* fix(core): point @experimental warnings at user code, not stdlib internals
Previously the wrappers installed by @experimental called warnings.warn
with a fixed stacklevel=3. ABCMeta inserts an extra abc.__new__ frame
when an experimental ABC is subclassed, so the warning landed inside
abc.py (or <frozen abc>:106 on modern CPython) instead of the user's
class Sub(...) line.
Resolve the user frame by walking inspect.currentframe(), skipping
frames whose module name is abc/functools/typing/contextlib (or
submodules), then emit via warnings.warn_explicit so the recorded
filename/lineno point at user code. Falls back to warnings.warn with
stacklevel=2 if no user frame is found. Module-name matching is used
because frozen stdlib modules report '<frozen abc>' as their filename.
Also install a one-line warnings.formatwarning specifically for
FeatureStageWarning so 'file:line: ExperimentalWarning: [ID] Name ...'
prints without the secondary source-snippet line. Other categories
delegate to the stdlib default formatter unchanged.
Added a regression test that subclasses an @experimental ABC inside
warnings.catch_warnings and asserts the recorded filename equals the
test file.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* fix(core): address review feedback on @experimental warning fix
- Make _install_feature_stage_formatter idempotent: tag the installed
formatter with a marker attribute and short-circuit re-installation,
so re-imports/reloads don't wrap the formatter on top of itself.
Also expose the previous formatter via __wrapped__ for restoration.
- Avoid leaking frame references in _resolve_user_frame: capture data
into plain locals inside try and del frame/candidate in finally,
per CPython's guidance on inspect.currentframe usage.
- Drop redundant _WARNED_FEATURES.clear() in the new ABC subclass test
(the autouse fixture already handles it).
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* changed query for foundry web search test
---------
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Enable instrumentation by default
* Update samples
* Optimization when span is not recording
* Address Copilot comments
* Revert uv.lock
* Add warning
* Formatting
* Fix mypy
* Add disable_instrumentation() with sticky user-intent semantics
Add a public disable_instrumentation() entry point so users can explicitly opt
out of Agent Framework telemetry, with a sticky-disable flag that makes the
user's intent "leading" — no framework code path (foundry's
configure_azure_monitor, configure_otel_providers, enable_instrumentation,
enable_sensitive_telemetry, or direct OBSERVABILITY_SETTINGS.enable_*
writes) can re-enable instrumentation until the user explicitly clears the
disable with enable_instrumentation(force=True) /
enable_sensitive_telemetry(force=True).
Also addresses the two remaining unresolved review threads on the PR:
1. test_observability_settings_defaults_instrumentation_true pins the new
"ENABLE_INSTRUMENTATION defaults to True when env unset" behavior.
2. test_enable_instrumentation_reads_env_sensitive_data restores coverage
for the post-import load_dotenv() fallback path.
Implementation:
- ObservabilitySettings.enable_instrumentation / enable_sensitive_data become
properties backed by _enable_*. While _user_disabled is True, the getters
return False and the setters drop True writes (defense in depth so third-
party writes can't subvert the disable).
- Public is_user_disabled read-only property lets integrations (e.g. foundry's
configure_azure_monitor) cheaply check the disable state without poking at
privates.
- enable_instrumentation() and enable_sensitive_telemetry() short-circuit with
an info log when disabled; gain a force=True kwarg that clears the disable.
- configure_otel_providers() still creates providers / exporters / views so a
later force-enable can use them, but logs an info message when called while
disabled.
- Foundry's FoundryChatClient.configure_azure_monitor and
FoundryAgent.configure_azure_monitor early-return when the user has
disabled, so Azure Monitor's global providers aren't installed unnecessarily.
Tests: 11 new tests covering default-on, env re-read at call time, sticky
behavior against each re-enable surface (enable_instrumentation,
enable_sensitive_telemetry, configure_otel_providers, direct attribute
writes), force=True override, re-arming the disable, and the __all__ export.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* docs: document disable_instrumentation() and force=True paths
Add a "Disabling instrumentation" section to the observability sample README
that walks through:
- The distinction between the ENABLE_INSTRUMENTATION env var (initial,
non-sticky) and disable_instrumentation() (process-wide, sticky).
- Why the sticky semantics matter: framework integrations like
FoundryChatClient.configure_azure_monitor() can call
enable_instrumentation() as part of their setup, and the user's opt-out
needs to win.
- All five surfaces guarded by the sticky disable (property reads, public
enable functions, configure_otel_providers, direct attribute writes,
is_user_disabled-aware integrations).
- The force=True escape hatch on both enable_instrumentation() and
enable_sensitive_telemetry().
- How third-party integrations should consult OBSERVABILITY_SETTINGS.is_user_disabled.
- The limits of the disable (does not tear down existing providers /
in-flight spans / third-party instrumentation, does not persist across
processes).
Cross-links the new section from the ENABLE_INSTRUMENTATION row in the env
vars table.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* docs: soften disable_instrumentation() overclaim about telemetry guarantees
Replace 'no telemetry will be emitted no matter what' (which is too strong,
since callers can still pass force=True or mutate private attributes) with
language framing the disable as a user-intent contract that library and
framework code is expected to honor: the framework actively short-circuits
the public enable paths, force=True and private-attribute writes are
acknowledged as out-of-contract escape hatches that integrations should
not use on the user's behalf.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* docs: correct observability Dependencies section
- opentelemetry-sdk is no longer a hard dependency; it is lazily imported by
create_resource(), create_metric_views(), and configure_otel_providers()
with a clear ImportError when missing. Day-to-day instrumentation works
with opentelemetry-api alone provided some other component configures the
global OpenTelemetry providers (Azure Monitor, an APM agent, application
bootstrap, etc.).
- opentelemetry-semantic-conventions-ai is no longer used anywhere in the
source; remove it from the listed dependencies.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* docs: replace stale observability migration guide with current PR's only relevant migration
The old guide documented the move away from setup_observability(otlp_endpoint=...)
which was an earlier-release API change unrelated to this PR and stale enough that
it's more confusing than helpful at this point. Replace it with a short note on the
single migration this PR introduces: callers of
enable_instrumentation(enable_sensitive_data=True) should switch to
enable_sensitive_telemetry(). Cross-link to the Disabling instrumentation section
for the rare 'force on without enabling sensitive data' use case where
enable_instrumentation() still applies.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
---------
Co-authored-by: Eduard van Valkenburg <eavanvalkenburg@users.noreply.github.com>
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Tao Chen
Yufeng HeandGitHub
Dineshsuriya DandGitHub
BaidarandGitHub